Fable 5: audit skills against Anthropic's new prompting guide — "too prescriptive" skills degrade output quality

## Context

Anthropic shipped Claude Fable 5 on June 9 together with an official prompting guide ([Prompting Claude Fable 5](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5)). One recommendation bears directly on gstack:

> Skills developed for prior models are often too prescriptive for Claude Fable 5 and can degrade output quality.

The guide's replacement pattern: give the model a goal, the reason behind it, explicit boundaries, and a way to verify its own work — rather than step-by-step scripts. There is currently no mention of Fable 5 anywhere in the repo (issues, commits, CHANGELOG), so opening this as a tracking issue.

## Measurements from a real install (v1.57.10.0, macOS)

- 54 skills registered in `~/.claude/skills`. Description metadata totals ~5.2 KB (~1.3k tokens) preloaded per session — this part is in good shape (terse one-liners).
- **49 of 54 SKILL.md bodies exceed the 500-line limit** recommended in Anthropic's [skill authoring best practices](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices). Top offenders by line count:
- `spec` — 2,275 lines
- `design-review` — 1,952
- `land-and-deploy` — 1,876
- `review` / `autoplan` — 1,810 each
- Every invocation pays the mandatory session preamble first (46 `gstack/bin` calls in `/codex` alone) — the ~3,000-4,200 token tax already tracked in #886.
- Usage concentration (my transcripts since Feb): 11 of 54 skills account for 100% of invocations.

## Relationship to existing work

The carve program is already moving in the direction the Fable 5 guide asks for:

- v1.54.0.0 — /ship carved into skeleton + on-demand sections (-59% always-loaded)
- v1.56.0.0 — Token-reduction Phase B
- v1.57.0.0 — carve-guard + cso/document-release/design-consultation carved
- #1993 — terse skill rendering honored in user installs (closes #886)

## Asks

1. **Adopt the Fable 5 guide as explicit acceptance criteria for carving** — not just "smaller", but "goal + rationale + boundaries + self-verification" instead of MUST-laden step scripts, deleting instructions that only existed to compensate for prior-model weaknesses.
2. **Extend carving to the remaining top offenders** (`spec`, `design-review`, `land-and-deploy`, `review`, `autoplan`).
3. **Consider a model-aware rendering profile** (e.g. document `explain_level: terse` as the recommended setting for Fable 5, or a dedicated profile) so prior-model scaffolding doesn't ship to models that the official guide says perform better without it.

Happy to share the measurement scripts or test candidate carves on Fable 5 against real workloads (the 11 skills above are my daily drivers).

Related: #886, #949, #766, #1993